PayPal has placed a conversational AI agent on top of its treasury operations, letting staff ask plain-language questions about cash positions, currency exposure and financial risk. According to a Deutsche Bank case study, the system covers $53 billion in corporate cash and customer funds that the team once tracked across 90 spreadsheets. It is a clear example of an agentic AI tool moving from pilot into daily use inside a regulated financial function.
What the Agent Actually Does
Instead of building a new report for every question, treasury staff can type a request and receive an answer through a chat interface. The case study describes queries such as exposure by jurisdiction over a chosen period, or the risk profile of the investment portfolio. The agent runs on a single platform that the overhaul consolidated, so its answers draw on the same underlying cash data rather than on separate reports that may disagree.
Balcioglu has framed the design choice in terms of proximity to the business. In the case study, he says the treasury group is "closer to the business compared to most other treasury departments," and the agent is meant to make that closeness practical by letting people with business questions get answers without waiting on a specialist to pull the data.
The rollout was not only a chatbot. Cash positioning moved onto the treasury management system across regions, and the spreadsheets were retired in the process. The team also manages pricing and trading for more than 160 currency pairs to support cross-border payment flows, which means the agent has to make sense of a wide range of exposures without a human rebuilding each view by hand.
The Numbers Behind the Rollout
The reported results are specific, though most of them come from PayPal and its partners. According to the case study:
- The spreadsheets were retired, and about 1,700 hours were saved each year.
- Automated cash visibility across roughly 1,000 accounts rose from 78% to nearly 100%.
- Integrating payments data into forecasting pushed cash-positioning accuracy above 99%, a figure PayPal presents as its own claim.
The 1,700-hour figure is also worth a closer look. Hours returned to treasury staff are capacity that can move from routine reconciliation into analysis, which is the kind of change finance leaders tend to measure first. Those numbers matter because treasury is a function where errors are expensive. A system that answers in seconds is only useful if the positions underneath it are correct. The case study ties the accuracy claim to the data integration work rather than to the chat layer, which is the right place to look. Enterprises evaluating similar tools should ask the same question: where does the answer come from, and how is it reconciled?
Companion Agents Are Still Being Built
PayPal has not finished the job. The case study says the treasury conversational agent is operational, a counterparty analysis agent is nearly complete, and a bank account opening and closing agent is still in the design phase. A PYMNTS report describes two companion agents, covering counterparty data and bank-account data, as still in development.
The sequencing is telling. The first agent answers questions about data that already exists. The planned agents are meant to handle more structured workflows, such as opening or closing bank accounts, which carry more risk if something goes wrong. That pattern, read first and act later, is common in finance, where audit trails and human sign-off matter as much as speed.
The same treasury group is also exploring other rails. In the case study, PayPal's treasurer Can Balcioglu says stablecoins "could complement existing payment rails," which suggests the company sees its agent work as one piece of a broader modernization of how money moves.
Why It Matters
For enterprises weighing agentic AI projects, the PayPal case shows what a credible first deployment can look like. It is bounded, it draws on one trusted data source, and it removes a large amount of manual report assembly. The value is measured in hours saved and visibility gained, not in autonomous decisions taken on the company's behalf.
Balcioglu said the approach "will define how treasury functions in the future." Whether that holds will depend on whether the companion agents meet the same accuracy bar once they touch counterparty and bank-account data. Those are the places where a mistake turns into a payment problem.
The broader lesson is about sequencing. Consolidating many spreadsheets into one governed platform is an operational change first, and the AI layer only pays off once the data underneath it is clean. Teams that skip that groundwork will find that a conversational interface simply makes inconsistent data easier to query and harder to trust.
